This action might not be possible to undo. Are you sure you want to continue?

# An Overview of Operational Modal Analysis: Major Development and Issues

Lingmi Zhang Nanjing University of Aeronautics & Astronautics, Rune Brincker Aalborg University, Palle Andersen Structural Vibration Solutions A/S, lmzae@nuaa.edu.cn China Denmark Denmark

Abstract

An overview of the major development of operational modal identification is presented. It includes four time domain approaches, i.e. NExT-type, ARMAV model-based, stochastic realization-based, stochastic subspace-based methods, and two frequency domain approaches, i.e. FDD-type and output-only LSCF-type methods. The internal relation with traditional modal identification using input/output measurements for EMA, as well as with classical and modern system identification are revealed. The major issues in OMA are also summarized, which cover full or partial references, multiple sensor setups, nonstationary excitation, data-driven vs covariance-driven SSI, structural mode sorting, bias or unbiased modal parameter estimation, as well as mode shape scaling.

1

Introduction

Modal analysis has been widely applied in vibration trouble shooting, structural dynamics modification, analytical model updating, optimal dynamic design, vibration control, as well as vibration-based structural health monitoring in aerospace, mechanical and civil engineering. Traditional experimental modal analysis (EMA) makes use of input (excitation) and output (response) measurements to estimate modal parameters, consisting of modal frequencies, damping ratios, mode shapes and modal participation factors. EMA has obtained substantial progress in the last three decades. Numerous modal identification algorithms, from Single-Input/Single-Output (SISO), Single-Input/Multi-Output (SIMO) to Multi-Input/Multi-Output (MIMO) techniques in Time Domain (TD), Frequency Domain (FD) and Spatial Domain (SD), have been developed. However, traditional EMA has some limitations, such as: 1). In traditional EMA, artificial excitation is normally conducted in order to measure Frequency Response Functions (FRFs), or Impulse Response Functions (IRFs). FRF or IRF would be very difficult or even impossible to be measured in the field test and/or for large structures; 2). Traditional EMA is normally conducted in the lab environment. However, in many industrial applications, the real operation condition may differ significantly from those applied in the lab testing; 3). Component, instead of complete system, is often tested in the lab environment, and boundary condition should be reasonably simulated. Since early 1990’s, operational modal analysis (OMA) has drawn great attention in civil engineering community with applications for off-shore platforms, buildings, towers, bridges, etc. OMA, also named as ambient, natural-excitation or output-only modal analysis, utilizes only response measurements of the structures in operational condition subjected to ambient or natural excitation to identify modal characteristics. OMA is also very attractive for aerospace and mechanical engineering due to many advantages, such as: (1) OMA is cheap and fast to conduct, no elaborate excitation equipment and boundary condition simulation are needed. Traditional modal testing is reduced to be response measurement; (2) Dynamic characteristics of the complete system,

Free decay response (FDR) can directly be measured either from transient excitation or sudden termination of board band random excitation. The NExT-type of operational modal identification procedures can also be called as two-stage TD modal identification. damping ratio and mode shape coefficient that is identical to the one of the corresponding structural mode. Correlation function (COR) can either be estimated directly from stochastic response via correlogram. or from Power Spectrum Density (PSD) via inverse FFT via periogram. as aforementioned traditional TD MIMO: estimation of the time response function (TRF) as the first stage. Actually. which makes use of correlation of the TRF data. followed in 1988. an MIMO version of ITD [5]. As a consequence. [11]. Correlation Function (COR) as well as Random Decrement signature (RDD) can all be expressed as modal superposition or modal decomposition. Each decaying sinusoid has a damped natural frequency. Natural Excitation Technique (NExT) was proposed for modal identification from output-only measurements in the case of natural excitation [1]. an improved PRCE called Improved Polyreference technique (IPCE). The time response function (TRF) can be estimated by the following techniques: Inverse FFT of the FRF estimation to obtain IRF for time domain EMA. an overview of the development of operational modal identification is presented and major issues in OMA are discussed. and then proved to be correlation function (COR) of the response. . Data correlation version of ERA.1 OMA in Time Domain. Eigensystem Realization Algorithm (ERA).e. In this paper. the closed-spaced or even repeated modes can easily be handled. adopted from system realization theory in linear system analysis. and can be computed through many ways from random response data [12]. and hence suitable for real world complex structures.instead of component. Therefore. Major Developments of OMA 2. NExT has shown that the COR can be expressed as a summation of decaying sinusoids. developed in 1982 [2] as an MIMO extension of the SIMO Least Squares Complex Exponential (LSCE) [3]. implemented in 1985 [6] To reduce the influence of noise contamination in the IRF data. structural control. and applied for modal identification in 1984 [4] Extended Ibrahim TD (EITD). and extraction modal parameters from TRF data as the second stage. Hence. NExT actually is an idea or principle that suggests using correlation function (COR) of the random response of the structure subjected to natural excitation. was developed in 1987[7]. (5) Operational modal identification with output-only measurements can be utilized not only for dynamic design. 1. the identification algorithm used for OMA must be MIMO-type. a summation of exponentially decayed sinusoids. [13]. Free Decay Response (FDR). NExT-type procedures In early 1990s. (4) All or part of measurement coordinates can be used as references. Random decrement signature (RDD) was proposed as free decay response in the beginning [10]. (3) The model characteristics under real loading will be linearized due to broad band random excitation. COR can therefore be employed as impulse response function (IRF) to estimate modal parameters. i. but vibration-based health monitoring and damage detection of the structures. There are three major TD MIMO algorithms have been widely utilized in EMA: Polyreference Complex Exponential (PRCE). major multi-input/multi-output ((MIMO) TD modal identification procedures developed in traditional EMA can be adopted for OMA. A Time Response Function (TRF) is proposed to represent these TD signature or features [9]. at much more representative working points can be obtained. as ERA/DC [8]. Impulse Response Function (IRF).

There are three steps for modal identification: (1) extraction of Φ and Z matrices. B R = α ∫ e Acσ Bc BcT e Acσ dσ ⋅ C T 0 ∞ Obviously.A number of algorithms in the PEM framework has been proposed.It should be noted that the NExT-type modal identification procedures adopted from EMA are all developed in the deterministic framework. PEM-ARMAV methods have applied to ambient modal identification in the middle of 1990’s [15].2 OMA in TD. The application of PEM to estimate an ARMA model results in a highly nonlinear optimization problem. (2) calculation of modal frequencies and damping ratios from Z matrix. (3) computation of modal participation factor matrix Γ. derived from first order equation of linear time-invariant (LTI) system in modal coordinates: H (t ) = Φe Λt Γ T or H k = ΦZ k Γ T . These algorithms identify the parameters of a model by minimizing the prediction errors. On the other hand. i. COR matrix is utilized. instead of IRF matrix. The COR matrix can be derived from stochastic state-space representation as R yy (t ) = Ce Act B R . ΓR = α ∫ e Λσ Γ T Γe Λσ dσ ⋅ Φ T 0 ∞ It is seen that only first two steps are required for estimation of modal frequencies and damping ratios and mode shapes. developed for EMA. in OMA the data utilized for modal parameter estimation are random response. x = ⎧ξ ⎫. However. instead of IRF. A = ⎡ ⎨ &⎬ c ⎢ −1 −1 ⎥ y (t ) = Cx (t ) y (t ) = cξ (t ) ⎩ξ ⎭ ⎣− M K − M D ⎦ In the modal decomposition formula. Corresponding to multiple natural excitations. modal decomposition of COR matrix. traditional ERA is based on the system decomposition of IRF matrices: H (t ) = Ce Act Bc Modal frequencies and damping ratios can be computed from the eignevalues of the system matrix Ac . multi-dimensional ARMA model. respectively. Φ. as an OMA procedure. etc. from which modal frequencies and damping ratios can be calculated. In the OMA case. Γ are mode shape matrix and modal participation factor matrix. 2. Z = e Λ∆t . However. The most important traditional system identification techniques are the Prediction-Error Method (PEM) [14]. t k = k∆t The formula is derived from differential equations of linear time-invariant system (LTI): 0 I ⎤ & & & Mξ&(t ) + Dξ (t ) + Kξ (t ) = bf (t ) or x(t ) = Ac x(t ) + Bc u (t ) . such as PRCE. EITD. No modal participation factor matrix can be obtained due to missing of input information. can be employed for OMA. mode shape and modal participation factor matrices can be calculated from the eigenvector matrix Φ plus output matrix c and input inference matrices b. In the OMA case. ARMA-type procedures A general auto-regression moving average (ARMA) model can also be employed for operational modal identification.. the difference in theoretical background and basic formulation should be notified. no modal participation factor matrix can be identified.e. system matrices Ac and C can be estimated as before via ERA algorithm. is utilized as basic formulation for modal identification: T T R yy (t ) = Φe Λt ΓR . Although the major traditional modal identification algorithms. ERA. which is stochastic process. Vector ARMA or ARMAV model should be applied. PRCE and EITD procedures are based on the modal decomposition of impulse response function (IRF). Λ is the eigenvalue matrix of Ac. Modal parameters can be computed from the ARMAV model by the coefficient matrices of the AR .

The LMS-ARMAV method overcomes some of the difficulties that have rendered ARMAV identification in use for mechanical structures. A Linear Multi-Stage (LMS) ARMAV method for effective OMA in the presence of noise is proposed recently [16]. . i. Stochastic Realization-based Procedures The ARMAV model-based methods are primarily used with so-called black-box model structures. for the modal identification. stochastic realization based procedures are often called as Covariance-driven Stochastic Subspace Identification (SSI) methods. The dimension of the AR coefficient matrices is equal to the number of inputs or references for PRCE. Making use of “past” output data as “instrument variable”. or “data-driven” approach. [22].e. Due to noise contamination. stochastic system realization was developed in 1970’s [23]. which is assumed uncorrelated to the input (white noise) residuals. System realization. Since eigenvalues of the system are calculated via SVD of IRF matrices (for EMA). The use of such models is quite cumbersome in multivariable case.3 OMA in Time Domain. System identification based on ARMA model is a non-linear process and should be implemented in iteration way. or MExT-type PRCE and EITD. a statespace model is preferred. 2. or “covariance-driven” approach. where a scheme for recovering the system matrices from impulse response function (IRF) is outlined [20]. In the modal community. Almost in parallel. belong to two-stage. It should be noted that covariance function (COV) is equal to correlation function (COR) for the zero-mean random process. It is worth noticing that PEM-ARMAV is one-stage. or covariance matrix (for OMA). and IVM-ARMAV methods. Refinements of the scheme are reported in 1974 and 1978. the original ARMA model turns to be AR model and its coefficient matrices are nothing but covariance (COV) matrices of the output data. It leads to factorization of Hankel covariance matrix. i. and number of output for EITD. especially for large dimension structures. However. For large scale system. The key feature of the stochastic system realization is the system decomposition of the COV matrix instead of IRF matrix in deterministic system realization. respectively. The non-linearity is caused by the MA parameters. The SVD-based system realization was firstly adapted for modal identification and named as Eigensystem Realization (ERA) in 1984. Instrument Variable Method (IVM) can be applied.e. Modal parameters can then be calculated by the eigenvalue decomposition of the companion matrix of the AR polynomial [17]∼ [19]. that means only the coefficient matrices of the AR polynomial are required. Two major drawbacks are inherent for PEM-ARMAV type OMA procedures: computational intensive and requirement of initial “guess” for the parameters to be identified. was developed way back to 1960’s. based on discretetime stochastic state-space equation. where COR is utilized for modal identification. recovering or identification of system matrices. Although derived in different way. A classical contribution in deterministic system realization by Ho & Kalman was published in 1966. For this purpose. introducing singular value decomposition (SVD) as a tool to reduce the noise inference in the IRF measurements [21]. It makes ARMAV-type procedures rather difficult to apply. As mentioned before that NExT-type ERA for OMA is actually one of the stochastic system realization approaches based on state-space model. and applied to modal identification in middle of 1980’s [24]. only modal parameters are interested. another important approach in the classical system identification. the final equations from the ARMAV model-based IVM are exactly the same as from NExT-type PRCE or EITD.polynomials. the system realization-based approaches can also be taken as subspace system identification method. actual model order should be defined much higher in practice.

2. ERA. As in the deterministic case.g. ordinary least squares (OLS). In PC method. The principal components of the resulting Hankel matrices HPC. not PC one. a new subspace-based state-space system identification (4SID) is developed in system and control engineering.r. 4). (2) The Canonical Correlation. The main procedure of the stochastic realization-based operational modal identification can be summarized as following four steps: 1). is a stochastic counterpart of the deterministic realization/identification algorithm. Computation of discrete-time stochastic system matrices via LS techniques. Making use of state predictor/estimate leads to Kalman filter for LTI system. computer simulation. as well practical application. In CVA and UPC methods. the other approach is to utilize observabily matrix and its shift to estimate system matrices. future response (CVA). CVA and UPC implementations of the stochastic realization-based OMA procedures. or mutual information w. etc. or Canonical Variant Analysis (CVA) method and (3) The Un-weighted Principal Component (UPC) method.e. The major difference between two . It is interested to note that UPC method. and can be expressed by so-called innovation state-space equation model: ˆ ˆ x ( k + 1) = Ax ( k ) + Ke( k ) ˆ y ( k ) = Cx ( k ) + e ( k ) Where the predicted state x ( k ) is called Kalman state. Since the state vector is not observable. Therefore. which offers numerically reliable and effective state-space model for complex dynamic system directly from measured data [26]. There are few possible numerical solutions to estimate the system matrix. total least squares (TLS) and partial least squares (PLS) techniques. its prediction (or estimate) should be utilized. Calculation of modal parameters from the system matrices. in the Gaussian case. HCVA and HUPC are. Instead of Hankel matrix shift. or retaining maximum reconstruction (prediction) efficiency for the output (UPC). reveals that no significant accuracy difference has been observed for PC. Stochastic Subspace-based Procedures In 1990’s.g.t. No non-linear search is required and computational complexity is therefore dramatically reduced compared to PEM-ARMAV or LMSARMAV procedures. 3). UPC is also called as Balanced Realization (BR). Estimation of covariance matrix from measure output data. weighted Hankel covariance matrix is adopted via different weighting [25]. K is Kalman Filter Gain and e(k) is the ˆ innovation. The physical explanation of the three methods in stochastic system realization is different. of course. which is corresponding to ERA. It can be proved that the optimal predictor/estimate of the state.There are three major methods for implementation of stochastic realization-based or covariancedriven SSI procedures: (1) the Principal Component (PC) method. in order to predict the system response. 2). Hankel covariance matrix is directly utilized for SVD. However. e. all different. to select partial states maximizing correlation w. output prediction (PC).r. i. block Hankel matrix and its shifted version can be used to estimate the system matrix.t.4 OMA in Time Domain. which is a zero-mean Gaussian white noise process. Stochastic Subspace Identification technique (SSI) has been followed making use of output-only measurements for a system subjected to stochastic excitation [27]. Its linear algebraic explanation is the orthogonal prediction of the state vector to the “past” output data vector. SVD of (weighted) Hankel covariance matrix to estimate observability and stochastic controllability matrices. e. It can be shown that UPC method performs balanced model reduction on minimum-phase model corresponding to the given covariance. is the condition mean of the state given all previous output measurements.

The advantages of SSI are that (1) it makes direct use of stochastic response data without estimation of covariance as first stage. G yy ( jω ) = H ( jω ) * G xx ( jω ) H ( jω ) T Where Gxx(jω). The main procedure of the SSI technique can be summarized as following four steps: 1) Computation of the projection of the row space of the “future” outputs on the row pace of the “past” outputs via robust numerical technique such as QR decomposition. but also for color noise. is a real Where rth scalar for white noise excitation. innovation. i. 2) Estimation of Kalman filter state via SVD of the above projection matrix. Gxx(jω) equals to constant. i. (2) the Principal Component (PC) method and (3) the Canonical Variant Analysis (CVA) method. Gyy(jω) are input and output PSD matrix. The PP technique gives reasonable modal estimates if the modes are well separated [29]. 4) Calculation of modal parameters as before. When all output measurements are taken as references. φ r . the stochastic subspace–based identification can be implemented in three methods depending on the choices of weighting matrices for projection matrix: (1) the Un-weighted Principal Component (UPC) method. which can be expressed as partial fractions form via poles λr and residues Rr N Rr Rr* H ( jω ) = ∑ ( + ) jω − λ* r =1 jω − λ r r Where Rr = φ r γ rT .e. H(jω) is the FRF matrix. then H(jω) are square matrix and γ r = φ r . Assuming input is white noise process. the modal decomposition of the output PSD matrix Gyy(jω) can be derived as modal decomposition: N ⎛ Ar ⎞ ArH Ar* ArT + + + G yy ( jω ) = ∑ ⎜ ⎜ jω − λ − jω − λ* jω − λ* − jω − λ ⎟ ⎟ r =1 ⎝ r r r r ⎠ pole λr = −σ r + jω dr . The datadriven SSI is belongs to one-stage modal identification category. 3) Estimation of the discrete-time system matrices via LS techniques. and that the two input/noise processes have been converted into one. and mode shapes can be obtained as a column of the PSD matrix at the corresponding damped natural frequency. 2.state-space representations is that the state vector is substituted by its prediction.e. γ r are mode shape and modal participation vector. corresponding rth residue Ar ≈ d rφ r*φ rT .5 OMA in Frequency Domain. Since SSI makes direct use of stochastic response data to identify modal parameters. PP is based on the fact that modal frequencies directly obtained from the PSD plot at the peak. respectively. FDD-type Procedures Frequency domain OMA methods are simply based on the formula of input and output power spectrum density (PSD) relationship for stochastic process: [28]. In the vicinity of a modal frequency the PSD can be approximated: 2d r φ rH = α rφ rφ rH G T ( jω ) ≈ φ r yy jω − λr ω →ω r Classical FD approach is Peak Picking technique (PP). it is also called data-driven SSI. SSI has soon been adopted for operational modal identification. respectively. (2) it cannot only be employed for white noise excitation. Similar to the stochastic realization-based approach. d r = γ rH G xx γ r . The main advantages compared to the TD techniques are that it has no bother with computational modes and .

The first generation of FDD can only estimate modal frequencies and mode shapes. the SV function can suitably be utilized as modal indication function (MIF). the data near the peak of the SV plot. the major FD modal identification approach is based on parametric transfer function model represented by rational fraction polynomial (RFP). The challenge is if we could develop a FD technique that has all the advantages but doesn’t have the disadvantages. estimated at discrete frequencies ω=ωi. The second generation of FDD. damping ratio estimation via half-power point is inaccurate or impossible. Therefore.e. From the free decay function of the S-DOF system. FSDD makes use of the singular vector. mode shapes can be obtained. It may cause bias error in damping estimation. but damping ratios [31]. A MLFD method was proposed making use of FRF measurements for . the rank of PSD matrix will be equal to the number of multiplicity of the modes. which is called as Enhanced FDD or EFDD has been followed for estimation of not only modal frequencies and mode shapes. PSD peak picking technique is inaccurate. H G yy ( jω i ) = U i S i U i When only rth mode is dominate at the modal frequency ωr. The accuracy of modal frequency estimation is limited to the frequency resolution of the PSD spectrum. The third generation of FDD. In most cases the enhanced PSD in the vicinity of a mode can be approximated as S-DOF system. for a complex structure. Since SVD has the ability of separating signal space from noise space. the modal frequency and the damping ratio can then calculated by the logarithmic decrement (Logdec) technique. which can leads inaccurate estimation of damping ratio by Logdec technique. which is approximation of correlation function of the SDOF system. the singular value data near the peak with corresponding singular vector having enough high MAC value are transferred back to time domain via inverse FFT. and closely spaced modes or even repeated modes can easily be detected. operational deflection shapes is obtained instead of real mode shapes. 2.e. The key of the second stage is conducting SVD of output PSD. The SISO. To do so. has been developed recently [32]. computed via SVD of output PSD with spatial measurements. From the corresponding singular vectors. i. beat phenomena would be encountered. to enhance the PSD. i.is much faster and simpler to use. SIMO and MIMO versions of RFP method were developed in 1980’s based on least squares solution [33]∼[35]. when dealing with closely spaced modes. Since only truncated data. Moreover. φ r = ur1 In the repeated mode case. In classical system identification. Modal frequencies can be located by the peaks of the SV plots. the modes can be indicated from SV plots. Moreover. The PSD matrix approximates to a rank one matrix as H G yy ( jω i ) = si u i1 u i1 ω i →ω r Compared to the previous PSD approximation formula. LSCF-type Procedures In the traditional EMA. are used for the inverse FTT to calculate approximate correlation function of the corresponding S-DOF system. A new FD technique named as Frequency Domain Decomposition (FDD) was proposed in 2000 [30]. PP technique is hardly applied for the structure with closely spaced modes. it is seen that the first singular vector at the ˆ rth resonance is an estimate of the rth mode shape. which is often encountered when dealing with real world complex structures. However. and therefore an S-DOF curve fitter can be adopted to estimate relevant modal frequency and damping ratio. Frequency-Spatial Domain Decomposition (FSDD). maximum likelihood (ML) estimators were developed to deal with noisy measurements [36].6 OMA in Frequency Domain.

multiple sensor setups are evoked.r. based on right matrix-fraction model. are only determined from observability matrix. To reduce the dimension of the involved matrices. these methods can be adopted for OMA based on modal decomposition of half PSD. It can be shown .modal identification in late 1990’s [37]. the formulations developed based on using all the output measurements as references. only limited of the sensors and/or data acquisition channels are available in the field response measurements. Nonstationary Excitation The output data will be nonstationary due to nonstationarity of the natural excitation in operational condition. only a subset of the outputs is required for references [41]. The SVD utilized in two approaches play a similar role in computing system matrices. 3.2 Multiple sensor setups In reality. However. and should be implemented as iteration process. In stochastic realization-based OMA or covariance-driven SSI.4 Data-driven vs Covariance-driven SSI There are clear similarities between covariance-driven and data-driven SSI. A least-squares complex frequency-domain (LSCF) estimation method was introduced to find initial values for the iterative MLFD method [37]. we can pick up any output measurements as the references for OMA via PRCE.1 Full or Partial References From NExT point of view. 3. computed from FFT of the COR with positive time lags via correlogram. has been developed recently [38]. as well as data-driven SSI. However. which is based on common-denominator model or scalar matrix-fraction description. to merge data from different setups will fail when response process is non-stationary for different records. A polyreference version of the LSCF method. In both covariance-driven and data-driven SSI. However. [40]! 3. and therefore the computational efforts. With p-LSCF. The aforementioned FD EMA methods are all based on modal decomposition of FRF matrix. and the excitation only affects the stochastic controllability matrix through the cross covariance matrix between state and output vectors. In parallel. there is theoretical question to be answered. 3. has two major shortcomings: (1) mode shapes and modal partition factor are difficult to obtain by reducing the residues to a rank-one matrix using the SVD. the above mentioned shortcomings can be eliminated. instead of FRF matrix [39]. Major Issues in OMA 3. EITD and ERA procedures as mentioned before. The first step of both approaches can be referred to as data reduction: the former via covariance estimation from Hankel data matrix and the latter via data vector projection by QR-factorization of the Hankel data matrix.3 Robust w. Therefore. Therefore the estimated covariance matrix with limited output data would not converge.t. LSCF. It is interested to note that normalization has extra advantage of smoothing out the nonstationarity in the data. so is the response data. (2) the closely spaced poles will erroneously show up as a single pole. It was found that these “initial values” yielded enough accurate modal parameters with much smaller computational effort. MLFD is a typical non-linear estimator. where the reference and real input point are not exactly the same. and therefore modal parameters. the system matrices. it is proved that approximate factorization of the covariance matrix still hold and the SSI algorithm can be applied to output data due to nonstationary excitation. When the excitation is nonstationary. Two approaches can not only be applied for data reduction but also data averaging/smoothing. A covariance matrix normalization technique has been developed to resolve this issue. However.

residue. the approach gives exact answers only when there is a full set of modes. which must be introduced to accommodate measurement noise. 2) Validation tools can be developed for data-driven SSI. covariance-driven algorithms can be fitted into the same framework of the data-driven SSI. since to scale a certain mode only that particular mode needs to be known. Recently a new approach based on repeated testing. are all based on sound theoretical background. can be reduced to solve a set of linear equations. On the other hand. including NExT-type of PRCE and EITD.g. 3) A spectrum formulation is available based on identified innovation state-space model. The main drawback is computation intensive and need initial “guess” for iterative search. and robustness for a truncated modal space has not been demonstrated. Theoretically.e. This approach seems more appealing. engineering practice has shown that little improvement can be obtained with rather heavy computational penalty. Also in health monitoring applications where damage is to be identified.that by an appropriate weighting for the Hankel covariance matrix. have limited effect. The p-LSCF technique usually results a clean stability diagram in wide frequency range. and therefore can better handle noisy measurements. residue. FD methods perform much better in structural mode determination. Time domain PEM and frequency domain MLE have the advantage of taking noise model into account. FDD techniques can almost eliminate spurious mode problem via modal indication function obtained from SVD. However. etc. including measurement noise. If the identified modal model is going to be used for structural response simulation or for structural modification.6 Bias and unbiased modal parameter estimation Most operational modal identification procedures. . 4) Modal decomposition of the total response can be made with data-driven SSI 3. There are a few advantages of data-driven SSI over covariance-driven ones: 1) The data-driven approach is numerically more robust due to its square root algorithm compare the matrix squared up in covariance-driven case. have been proposed [43]∼[45]. Some suggestion has been given in the literature for solving this problem [42]. stochastic realization-based and stochastic subspace-based identification methods. Over or underdetermination of structural modes results in inaccurate estimation of the modal parameters of the true structural modes. i. Mode Shape Scaling One of the critical issues is that OMA can not offer mode shape scaling due to lack of input information. the scaling factors might be important. modal indication functions and stability diagrams. IVM-ARMAV methods. Therefore..7. etc. instead of ordinary LS. leakage. as their counterparts in EMA. and having the capability of extracting unbiased modal parameters with confidence level. The tools developed until now. leakage. all the TD modal identification methods encounter a very serious and difficult problem in properly determination of model order and distinguishing structural modes from spurious or noise modes. 3. especially mode shapes and damping ratios. Even latest developed FD p-LSCF method can not avoid this bias issue. where mass changes are introduced at the points and the response is measured.5 Structural mode Sorting Most TD OMA methods. bias error can be reduced via employing TLS and PLS. e. non-linearity and un-modeled effects. then the scaling factors of the mode shapes must be known. However. Ordinary Least Squares (LS) solution is normally adopted. However. bias error could occur due to noise effects. nonlinearity and un-modeled effects. 3.

Feb. further theoretical proof based on system identification and stochastic process is still missing. and for structural health monitoring with vibration-based damage identification. equivalent to the one for FRF matrix. H. Moreover. 1992 [2] Vold. San Diego. 1979 .. instead of deterministic framework for EMA. The two-stage approach estimates covariance/correlation function (TD) or power spectrum density (FD) as first stage. Modal Testing Using Natural Excitation. It includes four time domain approaches.L. data-driven vs covariance-driven SSI. i. R. Most output-only version of system realization-based and subspace-based modal identification methods can be adopted from their input/output counterparts. Carne.. when correlogram is utilized to estimate COR and half PSD is obtained by using only the COR with positive time lag.790221. H. Since modal decomposition formula for output PSD matrix. The OMA version of ERA is actually one of the stochastic system realization algorithms. stochastic realization-based and stochastic sub-space approaches. Proc.P. 3-7. i. Parameter Estimation Techniques for Modal Analysis. The major issues in OMA are also summarized in the paper. Operational modal identification methods can also be classified as two-stage and one-stage approaches.e. USA. NExT-type. of the 10-th IMAC. such as PRCE. can be directly adapted for OMA from NExT point of view. FDD-type and output-only LSCF-type methods. SAE Paper No. D. Covariance-driven and data drive stochastic subspace identification (SSI) are typical two-stage and one stage method respectively. and two frequency domain approaches. However. It is interesting to notice that not only NExT-type procedures can clearly find their counterparts in traditional EMA. J.. PEM-ARMAV methods are typical one-stage approach. as well as Mode Shape Scaling. Therefore. CONCLUDING REMARKS The major development of the operational modal identification methods for OMA is presented. et. SAE Paper No. It is interested to notice that OMA version of PRCE and EITD can more properly derived from ARMAV model via classical instrument variable method. nonstationary excitation and response. NExT-type procedures are typical two-stage approach. Although engineering applications have shown that favorable modal parameters can be extracted under the nonwhite or/and even nonstationary natural excitation. the modal decomposition formula for output PSD looks exactly the same as the one for FRF matrix. A. Mode shape scaling issue needs to be further studied when OMA is intended to be applied for structural response simulation. for structural modification. G.G. most of FD modal identification techniques developed for traditional EMA can easily be adapted for OMA.. structural mode sorting. 820194. Major TD modal identification methods in traditional EMA. T. REFERENCES [1] James. it should be aware that the theoretical basis is different. multiple sensor setups.4. EITD and ERA. Accuracy issue would be more significant when dealing with many measurements with very noisy data from complex structures in operational condition. A Multi-Input Modal Estimation Algorithm for Mini-Computers.e. Nard. CA. bias or unbiased modal parameter estimation. can be derived. 1982 [3] Brown. which cover full or partial references. al. and then extract modal parameters from output COV/COR or PSD data. Stochastic framework should be considered in OMA. Lauffer.. et al. ARMAV model-based. except modal participation factor is replaced by reference vector.

USA.. April 1982 [13] Asmussen. Stochastic System Identification for Operational Modal Analysis: A Review. System Identification. 1(4). A. NASA/JPL Workshop on Identification and Control of Flexible Space Structures. Parametric Time-Domain Methods for the Identification of Vibrating Structures⎯A Critical Comparison and Assessment. ASME Transactions. Dept. Automatic Control. 1997 [16] Petsounis. Of IMAC XXI.1977. R. H.M. R. 123. S. “Random Decrement Technique for Modal Identification of Structures. J. Dec.. Yao. 1987 [8] Juang. Control-Theory and Advanced Technology. 1994. Automatic Control. 2001 [20] Ho. On the Two-stage Time Domain Modal Identification. 4(1).. J. Mechanical System and Signal Processing. IEEE Trans. Ph. 104... 1997 [14] Ljung.D. 1978 [23] Akaike. Approximate Linear Realizations of Given Dimension via Ho’s Algorithm. K. 1987 [15] Andersen. IEEE Trans.[4] Juang. “On-Line Failure Detection and Damping Measurements of Aerospace Structures by Random Decrement Signature.D. ASME Paper Number 85-DET-90..A. Pasadena. An Improved Time Domain Polyreference Method for Modal Identification. Identification of Civil Engineering Structures using Vector ARMAR Model. n6. -N. Fladung. L.-M. Investigation of Multiple Reference Ibrahim Time Domain Modal Parameter Estimation Technique. [19] Peeters. J. Mechanical System and Signal Processing.501-514. of Mechanical Design. Single Simple Modal Identification of a Nonstationary Stochastic Process. and Fuche. Proc. Brown. pp. Denmark. M. 1966 [21] Zeiger. An Eigensystem Realization Algorithm (ERA) for Modal Parameter Identification. 1974 [24] Benveniste. -X.. Allemang. J. Measurement and Control.. ASME J. 1986 [7] Zhang. Of 12th Asilomar Conf.L. L....J. Stochastic Theory of Minimal Realization. CA. M. pp. 1984 [5] Ibrahim. Thesis. et al. A.L. Proceedings of the IMAC. J. B. University of Cincinnati. and Pappa...J.. L. H. Modal Analysis based on the Random Decrement Technique. R..” Journal of Spacecraft and Rockets. 1977 [6] Fukuzono. And Wright. R. 14. [9] Zhang. S. 2001 [17] Leuridan. 2003 [10] Cole.. B. Theory for the User.8 pp. Aalborg University. Time Domain Parameter Identification Methods for Linear Modal Analysis: A Unifying Approach. Efficient Construction of Linear State Variable Models from Input/output Functions. 1973. H and McEwen. Stress. K. Denmark. D.. The Shock and Vibration Bulletin 47. W. 1985. of Mechanical and Industry Engineering. A Mathematical Basis for the Random Decrement Vibration Signature Analysis Technique. R. An Eigensystem Realization Algorithm Using Data Correlation (ERA/DC) for Modal Parameter Identification. J. and De Roeck.. G. R. Thesis.D. AC19. 1974 [22] Kung. [11] Ibrahim.. Palle. of Dynamic System. D. [12] Vandiver.C. Sept. and Kalman. -F.J. Vol. J.. No.. Ph.15. and Reliability in Design. A Method for the Direct Identification of Vibration parameters from the Free Response.A..” NASA CR-2205. 1985 . –N.C. and Fassois. Automatic Control. AC-30. S. V. System and Computers.J. IEEE Trans. Prentice Hall.-M. Modal Parameter Estimation: A Unified Matrix Polynomial Approach". Vo.. Regelungstechnik 14.R. Acoustics. [18] Allemang. K. Journal of Vibration.S.E. Cooper. J.. A New Identification and Model Reduction Algorithm via Singular Value Decomposition. 1988. on Circuits. Y. E. Proc. 696-700. & Mikulcik. Englewood Cliffs. J.. E. v. Thesis. Aalborg University. Brown.S. & Lu.

. Balanced Approximation of Stochastic System. & Piersiol. and De Moor.-M.G. 1996. IMAC XXI. Orlando (FL). 2002 [43] Parloo. Andersen. John ‘Wiley 7 Son. USA.. 1982 [34] Richardson M H. J. Leuven. 2005 [41] Peeters. In Proceedings of IMAC XXIII. B... Orlando (FL). IEEE Press and John Wiley & Son. 11(1).. Modal Identification from Ambient Response using Frequency Domain Decomposition. et. Rune Brincker and Canteli.. Leuven. F. Proc. D and Gunes. B. <Random Data. Mechanical System & Signal Processing. Reference-based Stochastic Subspace Identification for Outputonly Modal Analysis. Y.. B. [40] Peeters. In Proc.G.L. B. C. B. [30] Brincker. A. November.. Proceeding of 3rd IMAC. P. A Poly-reference Implementation of the Least-Squares Complex Frequency Domain Estimator. Implementation..[25] Arun. and De Roeck. Automatica 29 (3).J. USA. PolyMAX modal parameter estimation from operational data. USA. 16(5). et al. Brown. 2005 .. R. Applications. 2001. A. al. Y. 2003 [39] Peeters. [27] Van Overchee. and Kung. 1999 [42] Bernal. P.. J. Zhang. Matrix Analysis Application. P. Sensitivity-Based Operational Mode Shape Normalization. 1985 [35] Shih. R. Kissimmee. Subspace Algorithms for the Stochastic Identification Problem. K. [36] Pintelon.. IOMAC. 2005 [33] Richardson M H.. et al.S. of the International Modal Analysis Conference (IMAC) XX. L. E. 2000 [31] Brincker. [32] Zhang. Frequency-spatial Domain Decomposition Technique with Application to Operational Modal Analysis of Civil Engineering Structures. 1998 [38] Guillaume. Global Curve Fitting of Frequency Response Measurements Using the Rational Fraction Polynomial Method. SIAM J. Damage Localization in Output-Only systems: A Flexibility Based Approach. of the IMAC XV. and Schoukens. Proc. IMAC XXI. San Antonio. Belgium. C. Damping Estimation by Frequency Domain Decomposition. A Frequency Domain Global Parameter Estimation Method for Multiple Reference Frequency Response Measurements.. IMAC XXIII. Frequency Domain Maximum Likelihood Identification of Modal Parameters with Confidence Levels. & Tomas H. Allemang. M. April. USA. 13(6). 1986 [29] Ventura. Kluwer Academic Publishers. R. Florida. van & De Moor. D. 1993 [28] Bandat. L. S. 0f ISMA 23.. September 2004. In Proc. 2003 [45] Aenlle. Proc. R. IMAC XIX. of the 18th IMAC. Tsuei. 2002 [44] Brincker. February. Mechanical Systems and Signal Processing. USA. Analysis and Measurement Procedures>. Some Methods to Determine Scaled Mode Shapes in Natural Input Modal Analysis. 2001 [37] Guillaume. et al. al. Wang.. et. T. Kissemmee (FL). 1997. 1988.Y.-M. of ISMA 2004. USA. Belgium. F. and Anderson. R. Orlando. Kissimmee.. Ventura. ISBN 0-7923-9717-7. P. Subspace identification for linear systems – Theory. 2(4). 1990 [26] Overschee. System Identification: A Frequency Domain Approach.. Operational PolyMAX for estimating the dynamic properties of a stadium structure during a football game... E. Parameter Estimation from Frequency Response Measurements Using Rational Fraction Polynomials. C. “Structural Assessment by Modal Analysis in Western Canada”. Denmark. Mechanical Systems and Signal Processing. TX. Los Angeles (CA). New York. Proceeding of 1st IMAC....Y. Copenhagen. And Tamura. January.. USA. P. and P Andersen: A Way of Getting Scaled Mode Shapes in Output Only Modal Testing. February. B.L.